My rig: what can I run?
Pick your graphics card (or several) and system RAM, Mac or mini PC and see all 111 models: which fit, which run with part of the model in RAM, how fast they run and how much context they hold. Speed and context are calculated, not measured estimate
Your hardware
Identical cards, llama.cpp layer split: memory adds, speed stays at one card's. Why
With system RAM set, models too big for the card show whether they run with RAM offload: part of the model on the card, the rest in RAM, as LM Studio and Ollama do. It is much slower, and we don't estimate its speed yet. How
The whole system memory is counted. How much of it the GPU can be given depends on BIOS and OS settings, and the rest of the system needs some too, so fit is optimistic here.
Your numbers, not checked against a spec sheet. Total memory for all cards; with Apple, the unified memory total, of which we count the share macOS lets the GPU use by default (why).
What do these mean?
- Context
- How much text the model can keep in view at once: your prompt, any files you paste in, and the chat so far. It is counted in tokens (a token is a word or part of a word), so 8K is about 8,000 tokens. Longer context needs more memory.
- tok/s (tokens per second)
- Estimated generation speed: how fast the model writes its reply. Higher is faster. These are estimates from memory bandwidth, not benchmarks; measured MLX runs on Macs are labeled separately.
- Quantization and file size
- Models are shrunk ("quantized") to fewer bits per weight so they fit in less memory. Most files here are Q4_K_M, about 4 bits per weight; where a model uses another quant, its name is shown next to the size. A smaller file needs less memory and runs faster but can lose some quality.
- "Fits"
- Memory math only: the file, 1.5 GB of headroom and at least 4K tokens of context fit in the memory your GPU can use. Your app (llama.cpp, LM Studio, Ollama, MLX) still has to support the model. "Weights-only fit, context unknown" means we can't check the context part yet.
- RAM offload
- When a model is too big for your graphics card, LM Studio, Ollama and llama.cpp can keep part of it on the card and the rest in system RAM. It runs, but much slower, because RAM is far slower than the card's memory. Mixture-of-experts (MoE) models suffer least.
- Finding your hardware
- Windows: Task Manager (Ctrl+Shift+Esc) › Performance › GPU. The card's name is at the top right; "Dedicated GPU memory" is its VRAM. Linux (NVIDIA): run
nvidia-smi. Mac: Apple menu › About This Mac shows the chip (for example M4 Pro) and the memory. System RAM and its speed (Windows): Task Manager › Performance › Memory shows the size and "Speed" (for example 6000 MT/s); DDR5 runs at 4800 MT/s and up.
Loading…
Every model on this rig
How we got these numbers
"Fits (memory)" is memory math only: the model file (its quant is named under the size), headroom and at least 4K tokens of 8-bit context fit. "Weights fit, context unknown" means the file and headroom fit but we have no KV cache data for the model, so how much context fits is unknown: it is not a proven fit. Neither checks that your runtime supports the model yet. Badges read the low end of the tok/s estimate: Fast generation at 40+ tok/s, Very fast at 100+. That is generation speed at short context; prompt processing and the slowdown in long chats are not in it, so badges are left off Tight rows and uncalibrated hardware (how speed is estimated).
Estimates assume llama.cpp (which LM Studio and Ollama also run on), the Q4 files listed on each model page, generation speed (tg128) at short context, and a single user. How speed is estimated · how context is estimated · Mac memory limit · where specs come from. A model fits when its file + 1.5 GB headroom fits and at least 4K tokens of 8-bit context fit beside it; tight means under 8K tokens with the default FP16 cache. Hardware specs are first-party (linked per GPU on the GPU pages; the full source list is in methodology). Prices as of Nov 13, 2026.
RAM offload (PC with system RAM set): the card's memory plus your RAM less 8 GB kept for the system count as one pool for fit and context. The share on the card is given for a stated context: the longer the context, the smaller it gets, because each layer's KV cache travels with the layer. Speed is not estimated until at least 5 verified runs back it (1 so far); a reported run shows on the rows it matches and doesn't count toward that. Below a model's full window, the max context on an offload row is a ceiling that fills the memory, marked so. Your CPU, a laptop's lower power, and LM Studio's "offload experts to CPU" setting all change offload speed (how offload is worked out).
MoE experts in RAM (the checkbox; LM Studio's "offload experts to CPU", llama.cpp --cpu-moe): for a mixture-of-experts model we have expert data for, only the routed experts go to RAM; attention, shared experts, embeddings and the whole context cache stay on the card, so the share on the card doesn't change with context. With the box ticked it applies to every MoE model it can, even one that would fit the card alone (the row says so). When it can't apply (no expert data, the experts don't fit your RAM, or the rest doesn't fit the card) the row says which mode it was evaluated with instead, and why. The experts' size is their share of the model's parameters, from its Hugging Face config (details).
As an Amazon Associate I earn from qualifying purchases.